User’s perception based on the reviews in English

After cleaning the reviews in English, there were 3,421 valid reviews out of 3,424.

2,851 reviews were positive and 372 were negative. 198 reviews had 3-star ratings and are as well presented in the table below in “all reviews”.

There is a large discrepancy in the length of the reviews. The minimal length has only two characters while the longest has 955. The negative reviews were on average longer than the positive ones. However, the longest review was a positive one.

The table below presents a comparison of the positive and negative reviews and their length.

Table 16: Statistical parameters of the review length (in characters) in English

To select the optimal number of topics that could best describe the reviews, a varied number of topics was tested within the model: 30, 20, 10 and 5. The model with 5 topics showed plausible results concerning content of the reviews.

The most frequent topic according to the graph below is number 1; the four further topics seem to have similar size. The size of the topic corresponds to the prevalence in the overall reviews. The bigger the topic is, the more frequently it appears in the reviews. There is also one cluster of three topics that have similarly frequent terms characterizing them. The axis in the model does not have certain definitions but helps to visualize the difference/relationship between the topics though the distance between them. An example of the topic visualization with the LDAvis is presented below in Figure 15.

Parameter All reviews Positive reviews Negative reviews

Min. review length 2 2 2

Max. review length 955 955 386

Average review length 29 25 34

Median review length 13 13 51

Figure 15: Size of the topics for the reviews about the Plantix app in English visualized with LDAvis

Topic number 1 concentrates on the identification of the plant’s diseases function of the app and its quality. It is characterized by such terms as “app, plant, good, disease, crop, identify, problem, recognize, photo” etc. The terms that are used in topic 1 represent ca. 33% of the terms (tokens) used in the reviews. The percentage assigned to this topic means the proportion of the terms regarding this topic in the entire text of the reviews. The more important the term to the specific topic, the higher it is positioned in the bar diagram. The example of the bar diagram and the ranking of the terms in accordance with the estimated term frequency within a selected topic, which is number 1 in this case, is presented below in Figure 16.

Figure 16: The most frequent terms used in topic number 1 visualized with LDAvis

The topic cluster that combines topics number 2, 3 and 5 reveals the farmers’

estimation of the three different aspects connected with the app, where the most commonly used terms are “good, great, nice, awesome, helpful”.

Topic number 2 praises the team who created the app Plantix and their job.

“agriculture, knowledge, nature, gardening, environment” are the areas where the app has a positive impact according to topic number 3.

Topic number 5 contains such terms as “nice, super, good, excellent, bad”

addressed directly to the app. Topic number 4 shows the aspects of the

importance of the language used in the app. The following terms are characterized by this topic: “language, india, hindu, add, kannada, telugu, tamil” – all these are dialects in India and, apparently, the farmers would like to have the app in those languages, too.

Figure 17: The top-30 most frequent terms in the overall reviews visualized with LDAvis The graph above demonstrates the top 30 terms that appear most frequently in all reviews in English, independent of the topic. Among them the top five are “nice, good, app, super, farmers”. The frequency of the words can be seen on the bar graph below; the more often the word is used in the reviews, the longer the bar.

This means that, in general, most of the reviews are positive and that the farmers are satisfied with the app and that it apparently provides feasible advantages in the recognition of plant diseases, which supports hypothesis number one.

In a further step, the positive feedback was analyzed. To do so, reviews with 4 or 5 stars were filtered. Positive reviews comprised approximately 80% of all reviews in English. The share of positive reviews shows that the users were generally satisfied with the service the app delivered. To give the model a concrete shape, both 5 and 10 topics were tested. It was determined that 5 topics were best suited to summarize the positive reviews.

Figure 18: The most frequent terms found in topic number 1 in the positive reviews of the Plantix app in English, visualized with LDAvis

Topic number 1 is the largest of the 5 topics and contains 26.5% of tokens (terms) from the positive reviews; that is why the top 30 terms found there are

presented above in Figure 18. According to the top 30 most relevant terms in topic number 1, the users appreciate the app itself, the idea behind it, the knowledge they receive via the app, and the community that helps solve their problems. Since most of the customers using the app, according to the analytics platform sensortower.com, come from India (80%), the term “India” is also one of the most commonly used in the reviews.

Topic number 2 includes the top terms about the plant disease identification function and the language. The proportion of tokens in this topic is also more than 20% in the corpus of positive reviews. Topics number 3 and 4 have many positive descriptive terms such as “good, helpful, amazing and awesome”. In the topic number 3, those adjectives refer to the job or work that Plantix does. In the topic number 4, these terms are connected to the learning and information that farmers receive through the app. Topic number 5 contains many descriptive positive terms like “nice, good, superb, happy” that refer to the app . ¹⁸

According to the positive reviews analyses in English, the users appreciate the main function of the app that helps identify plant diseases, as well as the knowledge and community support that it additionally provides.

Although the positive reviews represent the largest part of the evaluation, there are still 20% of the users that had several issues with the app. The negative reviews received 1 or 2 stars from users. After trying different combinations of topics, 3, 4, 5 and 6 remained. The most suitable number of topics for describing the negative reviews was 6.

Topics number 1 and 2 contain the terms that were mentioned most frequently in the reviews in comparison to the others. There is a cluster of 3 topics, including those most often mentioned, which means the negative reviews have something in common.

As the graph below shows, the most frequent terms in topic number 1 concern not finding information about the plant or not identifying it correctly.

Figure 19: The most frequent terms in topic number 1 appearing in the negative reviews of the Plantix app in English visualized with LDAvis

In topic number 2, the most frequent terms were also about the problems with the plant and disease identification. Apparently, especially guava and mango were often not properly recognized by the app.

In topic number 3, the users were not satisfied with the solution suggested by the app. Topic number 4 was about language; the Kannada language was mentioned

especially frequently; apparently, some of the users understand it better than Hindu. The terms of topic number 5 were just generally describing that the app was not working, and the users were feeling that they had wasted their time. And finally, the last topic, number 6, concerned problems with the login and sign-in process.

In line with the negative reviews in English, it is possible to conclude that although the plant image recognition generally works in most of the cases, the algorithm is still not perfect. Some users were extremely frustrated whenever the plant was not properly identified. As a result, the solution was also not correctly recommended.

Language also plays very important role for the users. There is a trend that indicates that the Kannada language should be the next to be implemented.

According to the majority of the English reviews, the app is fulfilling its promises to the users, i.e. identifying the plants’ diseases and offering a correct solution.

However, the process is still not working perfectly; some users still have problems finding the right diagnosis. The most commonly used terms in the reviews were

“good, awesome, helpful, superb”, which positively describes the perceived usefulness of the technology. Since the present study was focused on Germany, in the following section, the reviews in German will also be analyzed and compared to the ones in English.